Body fat compartment determination by encoder-decoder convolutional neural network: application to amyotrophic...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Neurodegenerative Disease Research

Background:

• Body composition, particularly abdominal fat distribution (subcutaneous adipose tissue [SAT] and visceral adipose tissue [VAT]), is altered in neurodegenerative disorders like amyotrophic lateral sclerosis (ALS).
• Accurate and automated quantification of SAT and VAT from medical imaging is crucial for understanding disease mechanisms and evaluating therapeutic interventions.
• Current semi-automatic methods for fat segmentation are time-consuming and may introduce variability.

Purpose of the Study:

• To develop and validate an automated method for discriminating and quantifying human abdominal SAT and VAT using T1-weighted MRI.
• To employ encoder-decoder convolutional neural networks (CNNs), specifically a U-Net-like architecture, for automated fat segmentation.
• To apply the developed algorithm to a patient cohort with amyotrophic lateral sclerosis (ALS) and compare fat distribution with healthy controls.

Main Methods:

• A novel automated algorithm based on a U-Net-like CNN architecture was developed for segmenting abdominal SAT and VAT from T1-weighted MRI.
• The algorithm was trained and validated on a dataset of 155 participants (74 with ALS, 81 controls), with data split into training (50%), validation (6%), and testing (44%) sets.
• Performance was evaluated by comparing CNN-derived segmentation masks against a reference standard using Dice coefficients.

Main Results:

• The automated CNN approach achieved high accuracy in segmenting abdominal adipose tissue compartments, with Dice coefficients of 0.87 ± 0.04 for SAT and 0.64 ± 0.17 for VAT in controls.
• Similar performance was observed in the ALS group, with Dice coefficients of 0.87 ± 0.08 for SAT and 0.68 ± 0.15 for VAT.
• The study confirmed significantly increased VAT/SAT ratio in ALS patients compared to controls, consistent with previous findings.

Conclusions:

• The CNN-based U-Net architecture provides an efficient and automated method for segmenting abdominal SAT and VAT from MRI, significantly reducing data processing time.
• This automated approach offers unbiased analysis of body fat components, potentially serving as a valuable tool in neurodegenerative disorder research.
• The automated quantification of body fat parameters may emerge as a potential biomarker or secondary outcome measure in clinical trials for conditions like ALS.